Identification and clinical evaluation of diagnostic biomarkers for ischemic cardiomyopathy based on machine learning and transcriptomics

Figures

• 

  Figure 1. Work flowchart of this study.

• 

  Figure 2. Screening DEGs of ICM. (A) The volcano plot of DEGs is shown when comparing ICM and normal samples. Downregulated genes are represented by blue circles, and upregulated genes are represented by red circles. (B) The heatmap shows the top 25 differentially expressed genes of ICM.

• 

  Figure 3. Identification of the most relevant module genes via WGCNA in ICM. (A,B) Based on the results of scale independence and average connectivity, β=7 was chosen as the optimal soft threshold. (C) Gene co-expression modules are displayed in different colors below the gene tree. (D) Heatmap of the association between modules and ICM. The cyan module shows a strong positive correlation with ICM. The correlation coefficient and p-value are represented by the numbers in parentheses.

• 

  Figure 4. The functional enrichment and protein-protein interaction (PPI) analysis of these intersecting genes. (A) The Venn diagram shows that there are 12 intersecting genes in the DEGs of ICM and the WGCNA MEcyan module genes. (B,C) Gene ontology (GO) and Reactome Pathway Knowledgebase (Reactome) enrichment analysis of the 12 intersecting genes. (D) The PPI network shows the relationships between these genes.

• 

  Figure 5. Evaluation of the twelve machine learning models. (A) Feature importance across the twelve algorithms based on Root Mean Square Error (RMSE) loss. “Label” denotes the clinical grouping variable used as an internal baseline. Error bars represent the standard deviation of the RMSE loss following feature permutations. (B) Cumulative residual distribution of the models. ROC curves of the twelve models. The top three performing models (ADA, GBM, and XGB) are highlighted with thicker lines. (C) Exploratory ROC curves of the twelve machine learning models. Note that the internal AUC estimates may be optimistic due to information leakage during feature selection. The top three performing models (ADA, GBM, and XGB) are highlighted with thicker lines.

• 

  Figure 6. Construction of the nomogram. (A) The visual nomogram is used for diagnosing ICM. (B) The predictive accuracy of the nomogram model is assessed through the calibration curve. (C) DCA curve, used to evaluate the clinical applicability of the nomogram model. In the DCA plot, “All” assumes all patients are positive for the disease, “None” assumes no patients are positive, whereas “Model” represents the diagnostic strategy based on established gene signature.

• 

  Figure 7. Evaluation of diagnostic value. (A) ROC curves for SEPP1, CILP, and FRZB in the training dataset. (B) ROC curves for SEPP1, CILP, and FRZB in the external validation dataset.

• 

  Figure 8. Analysis of immune cell infiltration in ICM. (A) Comparative analysis using vioplot, showing the percentage of 22 different types of immune cells in the ICM group and the normal group. (B) Correlation analysis of 3 ICM-related biomarkers with various immune cells. (C–E) Serum protein concentrations of the key targets measured in peripheral blood from two groups using ELISA.*p < 0.05, **p < 0.01, ***p < 0.001.